Deep Learning-Powered Retinal Vein Occlusion Classification: Performance Benchmarking Across State-of-the-Art Neural Network Architectures Using a Novel Private Dataset

Retinal vein occlusion (RVO) is a leading global eye disease after diabetic retinopathy and can cause permanent blindness if not treated promptly. Developing and analyzing methods to understand RVO and ensure timely treatment is essential to prevent vision loss. In this paper, different models to analyze Retinal Vein Occlusion have been discussed, further classifying them into Central Retinal Vein Occlusion (CRVO), Branch Retinal Vein Occlusion (BRVO) and Normal. The raw retinal images were collected from S.M.C.H, Assam and further analysis was done on our own color fundus photographs (CFP’s) processed dataset. Few state-of-the-art (SOTA) models namely VGG-16, VGG-19, ResNet-18, ResNet-50 and Inception-v3 have been implemented and comparative analysis was performed using various evaluation metrics like accuracy, F1 score, precision and recall, out of which the Inception-v3 model outperformed all by achieving an accuracy of 95% with a low computational time of 867.86 seconds.